Data Use: New findings with old data using cluster analysis

Editor's note: The following article is adapted from a presentation made at the 1989 Sawtooth Software Conference. Tara Thomas is a senior market research analyst with Blue Cross and Blue Shield of Iowa.

For the past four years, Blue Cross and Blue Shield of Iowa (BCBSI) has been engaged in extensive consumer research. The purposes have been to uncover how consumers make choices in a complicated decision environment and how they evaluate the component features of insurance programs. Of particular interest has been the influence of service oriented activities on the choices that consumers make. The eventual goal is to determine to what extent choices can be altered by the inclusion or exclusion of service features in health insurance product designs.

During the summer of 1987, BCBSI and IMR Systems, Ltd. began a process that ended with a useful and informative perceptual mapping study of the image of BCBSI in the Iowa marketplace. The research would have been useful to us if the study had ended with that final report. But the true value of the research has been realized with continued analysis of this and other data sets, using two separate clustering methodologies.

The purpose of clustering these data sets was to investigate the possibility that market segments exist for health insurance products. As we recognized the possibility of segments within the data sets under investigation, we began to discuss the strategic implication of these results. Identification and description of the market segments would facilitate product design and distribution, and allow us to position and promote these products appropriately. In addition, understanding consumer choice patterns might eventually allow BCBSI to "custom design" products in a way that would enhance the composition of our insurance risk pools.

Research approach

During the summer and fall of 1988, BCBSI undertook a critical reexamination of five different data sets. The purposes of this analysis were to determine whether or not definable market segments exist and whether segment preferences are reflected in product choices. Four of these data sets were created through ACA (ACA System for Adaptive Conjoint Analysis), the fifth through APM (APM System for Adaptive Perceptual Mapping). Each represented a distinct set of respondents facing very different health insurance decisions. One ACA study examined only service features of health insurance, while the others included both service and insurance attributes.

To uncover possible market segments, the data were clustered using two different procedures: an attribute cluster approach using SPSS/PC+ and a respondent cluster approach using CCA (CCA System for Convergent Cluster Analysis). Though none of the original studies were designed as segmentation studies, the findings were comparable across all five. Across the data sets, two distinct segments emerge in the health insurance industry: one that prefers traditional health insurance products, and one that prefers the features of HMO health insurance products.

Findings and strategic implications

These two segments-the "Traditional" and HMO segments-are comparable across all five data sets. Quite frankly, this is not the result that we expected. In addition, several data sets exhibit what we believe to be initial indications of other emerging segments in the market. Our interpretation is that the market is in transition; new segments will emerge as consumers become exposed to wider choices in health insurance and gain better understanding of their decision environment.

Once the segments were identified using cluster analysis, the data sets were separated to conduct simulation analyses. The purpose of running the simulations was to determine which BCBSI product might appeal to different market segments. Not surprisingly, and quite reassuringly, those consumers who preferred traditional insurance were attracted to traditional products. And consumers attracted to HMO style insurance preferred HMO products. This indicates to us that consideration of product development for a segmented market was in order for BCBSI.

As a final step, the service features of products were excluded from a second set of simulation analyses. When these service features were excluded, consumer preferences changed- sometimes dramatically. A substantial shift in preference away from the HMO alternatives (which are grounded in high service and convenience) toward the traditional insurance products was apparent for all segments. This indicated that consumers often select their health insurance coverage (when they have a choice) based on how important service is to them; many consumers were willing to give up their choice of physician and other freedoms associated with traditional insurance to obtain enhanced service.

Study background and data bases

Our interest in conducting this health insurance segmentation study was initially raised by an article by Woodside, Nielsen, Walters, and Muller published in the Journal of Health Care Marketing. The authors identified four distinct and identifiable segments in the market for hospital services: the "Old-Fashioned," "Value Conscious," "Affluents," and "Professional Want-It-Alls." These four segments seek different benefits from the hospitals they prefer, and have characteristic demographic profiles. The study confirmed the contention of Kotler and Clarke that consumers seek different benefits from health care. Some seek "Quality," some "Service," some "Value," and others "Economy." Review of this research raised the question of whether similar segments existed for health insurance.

Our finding of two distinct and definable segments in the health care market has recently been supported by an article published in American Demographics. In April, 1989, Thomas and Sehnert reported on "The Dual Health-Care Market." In particular, the authors identified "Traditional" and "Modern" markets-analogous to the "Traditional" and "HMO" markets we identified in the Iowa marketplace.

Data set characteristics

Five different studies were conducted by BCBSI over the past four years: four conjoint studies and one perceptual mapping study. Two of the studies were conducted with random samples of Iowans; the remainder with client groups. Exhibit I illustrates the basic information about the sample composition and survey methodologies used in these studies. Each study was designed for a unique purpose, and the attributes were custom designed for the types of insurance products under consideration. While there are overlaps in the attributes and particular levels included in the studies, the similarities that we will note in segment composition are not simply an artifact of identical studies being conducted with very similar consumers.

Exhibit 1
Blue Cross and Blue Shield of Iowa Research Included in Segmentation Study

Service Unit Study
Sample: 201 adult Iowans covered by health insurance
Methodology: Adaptive Conjoint Analysis
Attributes: Service components of health insurance policies

Image Study
Sample: 403 adult Iowans covered by health insurance
Methodology: Adaptive Perceptual Mapping
Attributes: Image oriented, with some product characteristics

Client Study 1
Sample: I 1 9 employees of a mid-Iowa retailer
Methodology: Adaptive Conjoint Analysis
Attributes: Product and service characteristics of health insurance plans

Client Study 2
Sample: 88 employees of an eastern Iowa county
Methodology: Adaptive Conjoint Analysis
Attributes: Product and service characteristics of health insurance plans

Client Study 3
Sample: 400 employees of a central Iowa company
Methodology: Adaptive Conjoint Analysis
Attributes: Product and service characteristics of health insurance plans

Methodology
Two different clustering methodologies were used to analyze the underlying structure of the five data bases. The first one described uses the statistical procedures included in SPSS/PC+ to develop variable clusters and attendant respondent clusters. This multi-stage process begins with clustering of attribute levels and uses these attribute level clusters to define centroids that enable clusters of individuals to be formed and profiled. The second uses the CCA software developed by Sawtooth Software that clusters respondents based on full attribute-level response profiles.

Attribute-based clustering
The cluster analyses conducted using this procedure were conducted on the five data sets before CCA was commercially available. The methodology developed is a compromise design that allowed clustering of large data bases using SPSS/PC+ QUICK CLUSTER. The analysis steps used are as follows:

1. Utility values for individual attribute levels were examined to determine which were most important to the consumers included in a particular study. The most preferred level from every attribute was always included. Unimportant attribute levels were excluded from the clustering analysis.

2. The remaining attribute levels were clustered as variables using the CLUSTER routine included in SPSS/PC+, and the resulting cluster structure was examined for its plausibility and internal consistency. These variable clusters formed the basis of the consumer segments that were identified and profiled in succeeding steps.

3. CLUSTER analysis was repeated with subsets of the attribute levels until a reasonable and believable solution was found. Attribute levels that clustered "too closely" with another level of that same attribute were excluded at each step.

4. Once an understandable and believable cluster of attribute levels was reached, these clusters were used to define centroids for clustering individual respondents using QUICK CLUSTER. Centroids were initially defined using "above average" utility values for the attribute levels included in a cluster solution and average levels for all other attribute levels. Membership of individuals in these clusters was determined through the QUICK CLUSTER algorithm, and consumer profiles were prepared.

Respondent-based clusters
CCA was used to cluster respondents into market segments. To implement the CCA methodology, we used the suggestions offered by Sawtooth Software in the User's Manual. The data in the UTIL.PLS files were standardized and centered, and a mixed starting point strategy was utilized. No "hold-out" sample was used, as our approach to this research was exploratory in nature. The average reproducibility and pooled-F statistics were used to diagnose the results and select the best clustering solution. The tables of Reproducibility Due to Chance Alone in Appendix E of the CCA Users' Guide were used to test the null hypothesis that definable clusters do not exist these data sets.

Cluster analysis findings

THE SERVICE STUDY

This study was designed to isolate the "service" components of health insurance products-for example, claims processing time, personality of customer service reps, and claims filing requirements-from the "insurance" aspects. The objective behind applying cluster analysis techniques to this data base was to determine whether different groups of consumers respond differently to varying service packages. If this was true, "service bundling" could be used to attract consumers to health insurance options that are appropriate for them from an actuarial perspective.

The service study utilized conjoint analysis and included such attributes as "how easy it is to understand the policy," "how do insurance representative treat you," "how frequently are claims updates mailed," and "is the claim paid directly to the physician or not." Respondents were also asked whether they would be willing to pay a higher monthly premium for more preferred service packages. Not surprisingly, consumers as a whole prefer simple and understandable coverage, helpful and informed insurance reps, and "one contact" resolution of claims questions.

Consumers do not want to pay more for enhanced service packages, and having the doctor paid directly is an important convenience. Less emphasis is placed on how communication with the insurance company occurs.

Variable-based clusters
Clustering of variables results in three segments in this service study-the "Efficients," "Automatics," and the "Personals." The Efficients are the smallest segment in this sample, and prefer a policy that is easy to understand, with one contact to the head office to resolve claims, and clarity of communications.

The largest segment, the Automatics, desire very little interaction with the insurance company and are willing to pay more for convenience. They do not want to file their own claims, they expect quick payment, and would like a contact at their company who is knowledgeable about insurance.

The remaining respondents desire the personal touch in insurance products-thus the name Personals. These respondents are willing to visit the insurance company personally to resolve claims and want updates on the status of claims and deductibles. They would like a personal insurance representative. The Personals are more likely to buy their own insurance and have more experience with filing claims.

Market share simulations for these respondents support the finding that "more is better." Many respondents prefer the service advantages and convenience that are offered by HMO type insurance coverage.

Table 1
Market Share Simulations

Respondent-based clusters
CCA clusters of this data set revealed two segments of respondents with differing needs and desires for service bundles. The larger cluster values clarity and consistency in the services offered by insurers. These individuals value "knowing the rules" associated with resolving questions, and desire a system where claims questions are resolved through the mail via the main office. They want to know where they stand on a regular basis and are willing to pay more to obtain this kind of system. The response patterns of these individuals reveal that these customers presume that certain things will happen "automatically" in the system, such as claims being paid in 10 days or payment will be made directly to the physician. It is possible that these individuals have not experienced any difficulties in resolving claims, or that they have limited claims-filing experience.

The other cluster wants a simple system for interacting with the insurance carrier. They want a system where they have an individual to contact to resolve claims questions. They expect claims to be paid directly to the physician within 10 days of filing, and they do not want to file claims themselves. They are not interested in a system that involves indirect contact or increased costs.

Market share simulations were prepared for each of the two segments. The unusual finding is that all respondents prefer the high levels of service offered by HMOs, although the "HMO Segment" exhibits a stronger preference than the "Traditional Segment." The remaining shares of preference are evenly split between service packages typical of BCBSI plans and those of other commercial carriers.

Table 2
Market Share Simulations

THE IMAGE STUDY

This study was one of the earliest research efforts undertaken by Blue Cross and Blue Shield of Iowa. The objective was to understand the perceptions that Iowans have of BCBSI in the marketplace, compared to our many competitors. Perceptual mapping was used to develop visual images of the health insurance market for Iowa, and to show the perceived position of various companies within the market. The basic map of the marketplace is illustrated in Exhibit 2. Of note is the fact that the HMOs cluster (in terms of customer perceptions) along the left edge of the horizontal axis. The other commercial carriers included in the research are located centrally and to the right. It is clear from this visual presentation that consumers distinguish among the various health care alternatives available to them in the marketplace. The question is whether there are segments within this market with differing perceptions and accompanying product preferences. Clustering was done on respondent perceptions of health care alternatives.

Variable-based clusters
The clustering of variables identified three clusters in this perceptual mapping study. To form these clusters, only service and product attributes were included. "Imagery" attributes were excluded, as the primary interest was focused toward service and product design. Again, these clusters could be described as Efficients, Automatics, Personals.

The Efficients are the largest group in this segmentation strategy. These respondents want an insurance company that is helpful and trustworthy, pays claims promptly, handles problems fairly, and answers questions quickly. They also want the insurance company to provide good, accurate service.

Those consumers desiring "Personal" service want their insurance company to be friendly, caring, and flexible in dealing with claims problems or questions. They do not want any surprises from the insurance company. These respondents represent the smallest group in the sample and are less likely to receive insurance coverage from their employers. The last group, the Automatics, represent the remainder of the sample, and are only concerned about one characteristic: coverage where there are "no claims forms to send in."

Respondent-based clusters
When CCA was applied to this data, two stable clusters of respondents were identified. The smaller cluster includes individuals who want a traditional insurance carrier-much like BCBSI or the other commercial carriers. These respondents are older, with lower incomes than others who participated in this research. Demographically, these respondents are similar to those described by Thomas and Sehnert. The perceptual map for these respondents (Exhibit 3) illustrates their view of the insurance marketplace. The horizontal dimension represents the traditional characteristics of insurance. The vertical dimension represents "service and human" characteristics.

The second cluster includes respondents who prefer the convenience of a "non-traditional" approach to health care. Their attributes of importance: no claims forms to send in, no annual deductible, community oriented, innovative, and nonprofit. The horizontal dimension of the perceptual map for these respondents (Exhibit 4) is characterized by typical HMO characteristics. The vertical dimension represents traditional insurance.

THE CLIENT STUDIES

The three client studies were designed to explore the preferences of consumers who are faced with multiple health insurance alternatives. The respondents in all three studies work for companies that either have implemented or will be implementing insurance options where the employee has a choice from among several insurance alternatives. Typically, these choices consist of traditional comprehensive major medical (CMM) insurance programs, managed care options like preferred provider organizations (PPO), and HMOs.

Of interest to insurers is how consumers will make their product selection: Will the final selections of consumers be based on level of coverage, style of health care coverage, or anticipated patterns of health care utilizations? And are there segments of consumers with clearly definable decision styles who will be attracted to particular products? Although we typically use these studies in a consulting role to help employers understand the needs and desires of their employees, the information gathered from these studies is very interesting to BCBSI from a product development and design standpoint.

These three studies are conjoint studies where the attributes are defined as components (both service and insurance features) of health insurance alternatives that the company is considering making available to their employees. The attributes can be bundled to create product configurations that match the products currently offered in the marketplace. The respondents in two of these studies have not experienced a choice of health insurance options before, and are therefore naive about implications of the choices they might make. The third group of respondents have experienced making choices, and should understand the implications of their choices.

Reanalysis of all three data sets indicates the presence of two or three segments in the data. One of the segments in each case is an HMO segment-appreciative of the convenience and full coverage of HMOs. The second segment is a "Free Choice" segment that prefers to have no limitation on the choice of physician. These respondents prefer a more traditional approach to the provision of health insurance. The third segments that exists in the data bases is harder to explain and understand. We hypothesize that this segment may consist of individuals who cannot integrate the complex information presented to them in a multiple-choice environment, and whose pattern of product and service preferences is therefore not as clear as for the other two segments.

In this report, the results for only one of the three data sets will be reported. All three data sets lead to similar conclusions, and the patterns of preferences and market shares are likewise similar across the data. The data that will be reported are for the 400 respondents of a mid-Iowa corporation.

These respondents have made health insurance choices in the past, and are familiar with the differences in coverage offered by the alternatives available to them.

Variable-based clusters
Two strong clusters were identified from the variable-based clustering methodology. The HMO cluster accounts for the largest proportion of the respondents. The preferences of this cluster are typical of the coverages offered by HMO products: All charges for all services are covered for participating physicians and pre-approval is necessary for some procedures and hospitalizations.

The other cluster values the freedom to choose their own physicians. This Free Choice cluster tends to be better educated and married. The product preferences exhibited by this group include the ability to choose any doctor, no annual out-of-pocket expenses, nationwide coverage, and no pre-approval requirements for treatment.

Two sets of market share simulations were run for each segment. The first simulation includes all of the service and product attributes included in the study. The second simulation excludes the service attributes. With all attributes included, consumers in both clusters prefer HMO alternatives. When the service characteristics are excluded, however, there are significant preference shifts toward the CMM and PPO care alternatives. Tables 3 and 4 illustrate these preferences.

Table 3
Market Share Simulations
All Attributes Included

Table 4
Market Share Simulations
Service Attributes Excluded

Respondent-based clusters
CCA again identified two main market segments: the Free Choice Segment and the HMO segment. The Free Choice segment includes respondents who prefer those characteristics normally associated with CMM and PPO product alternatives. These include the freedom to chose a physician, limited annual deductibles and out-of-pocket costs, and require that the consumer file all medical claims.

The HMO segment desires full coverage for all services, even if this means limitations on their choice of physicians. They also like automatic claims filing and the no annual out-of-pocket cost provisions of HMO coverage. These younger consumers have higher educational attainment than others included in the research.

The market simulations for these two segments mirror the pattern found with the variable based-clusters. Again consumers switch from HMO-style coverage when service attributes are excluded from the simulations.

Table 5
Market Share Simulations
All Attributes Included

Table 6
Market Share Simulations
Service Attributes Excluded

A market in transition
In this data set there is considerable evidence of a market in transition. Our belief is that the market is evolving, and that the impetus for this evolution comes from the consumers themselves. The development of alternative health insurance coverage systems such as HMOs and PPOs has not only revolutionized the health insurance industry, but the thinking of the consumer as well.

In this client study, two clear market segments emerged: a segment that strongly desires freedom of choice and a segment that prefers an HMO alternative. Diagnostically, however, the five-cluster variable-based cluster is as plausible as the three cluster respondent-based cluster. While difficult to summarize, careful and repeated application of these methods leaves us with the opinion that these additional segments are not ephemeral or transitory, but are very real indicators of an evolving market.

Variable-based clusters
The five-cluster solution here is as plausible as the strong two-cluster solution. It would be easy to dismiss this solution, except for the fact that the cluster sizes are large enough to be considered significant, and the demographic compositions of the clusters differ. It is possible that there are customers in this market who are not satisfied with the health insurance coverage decisions they have made in the past. This is best illustrated by Cluster Two, where the grouping of attribute levels does not make sense. This cluster is willing to pay $20 more a month for a program that requires pre-approval. This could indicate that they are willing to pay more while giving up some of their decision making freedom.

This combination of attributes may indicate that they believe they are not currently getting the kind of coverage they might prefer. Table 7 breaks out the attribute levels that dominate for each of these five clusters.

Table 7
Five Cluster Solution

Cluster One (19 % of the sample)
Full payment for certain doctors
All services covered
Deductibles for ER and physician visits

Cluster Two (21 % of the sample)
Pay $20 more per month
Pre-approval required

Cluster Three (21 % of the sample)
Full payment for any doctor
Pre-approval not required
No annual out-of-pocket expenses

Cluster Four (25% of the sample)
Claims filed automatically
Pay $40 more per month

Cluster Five (I 5% of the sample)
Pays at 80% for any doctor
Pays at 90% for certain doctors
Unlimited nationwide coverage
Pre-approval sometimes required

Respondent-based cluster
The CCA clustering methodology identified three viable clusters. The HMO and Free Choice clusters are similar to those described before. The third cluster, which is the smallest of the three, is more difficult to interpret. This group of respondents values the following characteristics of health insurance:

Pay $40 more a month
Pre-approval always required
Limited annual deductibles
Must file all claims
Regular updates on claims status
Limited annual out-of-pocket costs

Demographically, this third cluster differs from the sample in one way. This group of respondents has significantly lower educational attainment than other respondents. It is likely that this group is trying to avoid catastrophic costs often associated with health care. Another possible explanation is that these respondents are not satisfied with health insurance decisions they have made in the past, but do not know how to integrate complex coverage information to make a better or wiser decision.

From a technical standpoint, this market segment is very stable as the number of clusters increases from three to four and more. CCA output includes a "switching matrix" that details how many respondents switch from one cluster to another as the number of clusters increases. Very little "switching away" from this cluster occurs as the number of clusters increases.

Conclusions

There is considerable evidence in the recently completed segmentation work that the health insurance marketplace is in transition. The belief of the primary researchers is that the market is evolving, and that the impetus for this evolution comes from the consumers themselves. The development of alternative health insurance coverage systems such as HMOs and PPOs has not only revolutionized the health insurance industry, but the thinking of the consumer of health insurance as well.

The five data sets examined by BCBSI give clear indications of two strong consumer segments in the market: a traditional insurance segment and an HMO segment. These two segments are present and stable across the data sets. Several of the data sets, however, exhibit less stable, but potentially significant additional segments. The preferences of these segments are not always clear, but their presence cannot be denied.

The diagnostics for the client study reported indicate that the five-cluster variable solution may be worth examining. Although the two-cluster solution is the strongest, and clearly optimal, the five-cluster solution indicates the presence of two or three market niches whose product and service expectations may not be well met by either traditional or HMO products. The question is whether these consumers are slipping through a gap somewhere, and whether they might be more likely than others to change their coverage.

While they are difficult to summarize, careful and repeated examination of these two data sets leaves us with the opinion that the additional segments observed in the data are not ephemeral or transitory in nature, but are very real indicators of an evolving market.

The existence of distinct and definable groups of consumers with differing preferences for levels and forms of service raises an opportunity for BCBSI. There is the potential to integrate service into future product development and research in a way that enhances the probability that individual consumers will be attracted to the product that is most appropriate for them from an actuarial perspective. This research demonstrates that bundling of product and service features can be used to develop optimal products.

Part of this research was conducted on behalf of Blue Cross and Blue Shield of Iowa while the author was senior research associate at IMR Systems, Des Moines. The assistance of Mary Ellen Burr in programming and analyzing these data sets is gratefully acknowledged.

References
Kotler, P. and Clarke, R. (1987). "Marketing for Health Care Organizations." Englewood Cliffs, N.J.: Prentice Hall, Inc.

Thomas, R.K. and Sehnert, W.F. (1989, April). "The Dual Health-Care Market." American Demographics, pp. 46-47.

Woodside, A.G., Nielsen, R.L., Walters, F. and Muller, G.D. (1988). "Preference Segmentation of Health Care Services: The Old-Fashioneds, Value Conscious, Affluents, and Professional Want-it-Alls." Journal of Health Care Marketing, 2 (2), p.14.